EFITA 2003 Conference 5-9.

July 2003, Debrecen, Hungary

823
GREENHOUSE CLIMATE MODELS: AN OVERVIEW
Jos Boaventura Cunha, jboavent@utad.pt
UTAD Universidade de Trs-os-Montes e Alto Douro, Dep. Engenharias
CETAV Centro de Estudos e de Tecnologias do Ambiente e Vida
5001-911 Vila Real, Portugal



Abstract: Greenhouse climate and crop models are essential for improving environmental
management and control efficiencies. In this paper, are described several types of models that could be
used to simulate and predict the greenhouse environment, as well as the tuning methods to compute
their parameters. This study focuses on the dynamical behaviours of the inside air temperature,
humidity and carbon dioxide concentration models and their domains of application. Linear and non-
linear models will be covered, focusing on issues such as: physical models, black-box models, and
neural networks models. Several experiments will be presented to illustrate the performance of each
model in the simulation and prediction of the greenhouse climate. The models are described as
functions of the outside climate, the control actions performed and the transpiration and
photosynthesis responses of the plants. The data used to compute the simulation models were acquired
in an experimental greenhouse using a sampling time interval of 1 minute. The greenhouse is
automated with several actuators and sensors that are connected to an acquisition and control system
based on a personal computer.

Keywords: Crop models, Greenhouse climate models, Neural networks, System identification


1. Introduction
To improve the management and control of a greenhouse climate, an adaptive climate
control strategy must be used to compute the optimal control signals used for a defined cost
performance function. The adaptive PID Proportional Integral Derivative controller
structure showed in figure 1 could be used for this purpose. The adaptive controller needs the
use of greenhouse climate models to predict future outputs based on past and current inputs,
the expected control actions and the predicted weather inputs. An optimiser module computes
the proportional, integral and derivative gains, in order to minimize or maximize the specified
cost function according to the constraints imposed. Afterwards, these parameters are sent to
the PID closed controlled loop.
EFITA 2003 Conference 5-9. July 2003, Debrecen, Hungary
824
Greenhouse Greenhouse
model model
Opt i mi zer Opt i mi zer Const raint s Const raint s
Weat her Weat her
model model

Past I nputs
and Out put s
Predicted weather inputs
Expected
control
input s
Predicted
errors
Predict ed
Out put s
+
_ Reference
t raj ect ories
PI D PI D cont r ol cont r ol
al gor i t hm al gor i t hm
Cost Funct ion Cost Funct ion
PI D paramet ers

Figure 1. Structure of the climate adaptive PID controller
Normally, the optimisation is performed to maximise a cost function which has a
positive term related with the expected crop economic value and a negative term related with
the operation costs of the climate conditioning equipment. Also, constraints for the physical
and physiological values of the actuators, the environment and the crop are applied. This
paper describes possible implementations for the models used by the controller.
2. Greenhouse climate models
Simulation models to describe the dynamic behaviour of the air temperature and
humidity and dioxide concentration inside the greenhouses have been published in several
studies. These models could be based on energy and mass flows equations (Boulard et al.,
1993, Bot, 1991), or derived by using a system identification approach using linear and non-
linear techniques, such as the recursive least squares algorithms and neural networks to tune
the parametric models (Boaventura Cunha et al., 2000; J. P. Coelho et. al, 2002).

2.1. Physical based Models
Physical models describe the flow and mass transfers generated by the differences in
energy and mass content between the inside and outside air, or by the control or exogenous
energy and mass inputs, as showed in equations 1 and 2 (Bakker et.al., 1994).

( )
, ,
1
i
i n h out h h
aph
dT
q q p
dt C
+ (1)

( )
, ,
1
m
i n m out m m
dc
q q p
dt V
+ (2)
where T
ag
is the air temperature, C
aph
the thermal capacity, q
in,h
and q
out,h
the energy
inflow and outflow, p
h
the energy production per unit of time, c
m
the mass concentration, q
in,m
and q
out,m
the mass inflow and outflow and p
m
the mass produced per unit of time referred to
the greenhouse volume V(m
3
).
In the previous equations, the transport mechanisms for conduction, convection and
radiation are implicit. For instance, the heat flux from inside to outside due to ventilation,
which is a term of q
out,h
in equation1, is described by the following equation:

,
( )
vent h v p i o
q q c T T (3)
EFITA 2003 Conference 5-9. July 2003, Debrecen, Hungary
825
where q
v
is the volumetric flux trough the windows, c
p
the volumetric specific heat,
T
i
the inside temperature and T
o
the outside temperature.
The drawback of this methodology is that the development of these models are
difficult to tune in practice, since they use a large number of parameters and physical
variables. Moreover, when properly tuned, they can only provide good predictions over short
future time horizons, since the greenhouse-crop system is time variant.

2.2. Black-box Linear Parametric Models
This method is based on experimentation where the input u and output y signals from
the system to be identified, figure 2, are recorded and subjected to data analysis in order to
infer a model. This procedure is known as system identification (Ljung, 1987). In this case,
the output signal vector y is formed by the measurements of the inside air temperature and
humidity, T
i
and RH
i
. These models must be related to the external influences of the outside
weather conditions, as well to the control signals performed on the greenhouse actuator
equipment.
Gr eenhouse Gr eenhouse
model model
ee((kT kT))
noi se noi se
Gr eenhouse Gr eenhouse
syst em syst em

y = [ T
i
Rh
i
]
y = [ T
i
Rh
i
]

+
_
u = [ T
o
Rad, u
vent
. . .]
Gr eenhouse Gr eenhouse
model model
ee((kT kT))
noi se noi se
Gr eenhouse Gr eenhouse
syst em syst em

y = [ T
i
Rh
i
]
y = [ T
i
Rh
i
]

+
_
u = [ T
o
Rad, u
vent
. . .]

Figure 2. System and model representation
In this figure, e, y and denote a noise signal, the model output and the prediction
error. The model inputs u are formed by the measurements of the outside temperature T
o
,
Solar irradiation Rad, ventilator and heater control signals u
ven
t and u
heater
, among other
relevant measurements.
The system identification procedure could be developed assuming that the greenhouse
climate can be described as a linear system around a particular operating point. In this way,
linear parametric ARX models can be employed to describe the dynamics of the greenhouse
climate system (Boaventura Cunha et al., 1997; 2000, Ljung, 1987). Previous work has shown
that the second-order ARX model equations 4 and 5 describe the dynamics of the air
temperature and relative humidity well.

( )
( ) ( )
( )
1,t 2,t 3,t 4,t
1 2
1,t 2,t

1 ( ) ( )
( ) ( )
i o
is
pipe i heat
i o vent
T kT T kT
B B B B Rad kT
T kT
a q a q T T u kT
T T u kT

1
1
1
1 ]

1
+ +
1
1

]
(4)

( )
( )
( ) ( )
1, 2, 3, 4,
1 2
1, 2,
( ) ( )

1
( ) ( )
i o vent
i h h h h
is
h h i o
pipe i heat
Rh Rh u kT
T kT B B B B
Rh kT
a q a q Rh kT Rh kT
T T u kT

1
1
1
1 ]

1
+ +
1
1
]
(5)

where: a
i
denotes the denominator parameters of the transfer functions, B
i
the
polynomials in the delay operator, T
o
the outside air temperature, Rad the outside solar
EFITA 2003 Conference 5-9. July 2003, Debrecen, Hungary
826
radiation, T
i
and Rh
i
the measured inside air temperature and humidity, T
pipe
the temperature
of the heating pipes and u
ven
t, u
heat
the ventilation and heating inputs.
However, the simulations computed with this second approach, for sets of data that
were not used to compute the models parameters, are more sensitive to mismatches compared
to the physical models. This is due to the fact that the ARX black-box models are a great
simplification of the entire system. Therefore, the parameters are time- varying. As an
example, the process of solar irradiation conversion to heat varies throughout the day and the
year, since the Sun elevation and the optical properties of the cover varies in time.
To overcome this difficulty, recursive estimation algorithms must be implemented to
compute the time-varying parameters of the transfer functions of equations 4 and 5, (Astrm
et al., 1989). Since the parameters are slow time- variant and the signals have no excitation for
long periods of time, an estimator that forgets the information only in the directions in which
new information is gathered must be used. The following recursive least squares algorithm
with exponential forgetting, equations 6.1 to 6.3 can be employed with this objective,
(Boaventura Cunha et al., 2000; Salgado et al., 1988),

( ) ( 1) ( )( ( ) ( ) ( 1))
T
k k K k y k k k + (6.1)

( 1) ( )
( )
( ) ( ) ( 1) ( )(1 ( ) ( ))
T
P k k
K k
k k P k k k k

+
(6.2)

1 1
( 1) ( ) ( ) ( 1)
( ) ( 1)
( ( ) ( )) ( ) ( 1) ( )
T
T
P k k k P k
P k P k
k k k P k k





+
(6.3)
where denotes the estimated model parameters (coefficients of polynomials A and B)
and K and P are the gain and covariance matrixes. For this estimator, the estimations
converge to values such that P(k)=I, where typically 0.0001<<0.001, and (k) and (k) are
the regression vector and the estimate of the variance of the residuals.


2.3. Black-box Non-linear Parametric Models
Artificial neural networks are collections of mathematical models that reproduce some
of the observed properties of biological nervous systems. The key element of the ANN is the
structure of the information processing system. This system is composed of a large number of
highly interconnected processing elements that are analogous to neurons and are coupled
together with weighted connections that are analogous to synapses.
Non-linear autoregressive models are potentially more powerful than linear ones
because they can model more complex underlying characteristics of the data. There are a
broad number of ANNs topologies. Among the most widespread are feedforward networks. In
this paper, a multilayer perceptrons (MLP) network with a hyperbolic tangent (tanh)
activation function is used. These types of structure have proved to be universal aproximators
(Hornik et al., 1989). This means that they can approximate any reasonable function
f
with a
subjective accuracy given by:


1 1
( ) , 1...
k n
jl ij i j l
j i
f u v w u l m

_
_



,
,

(7)
where

is the tanh function, k is the number of hidden units,

jl
v
and
ij
w
are weights,
i

are biases and

u
the data vector.
EFITA 2003 Conference 5-9. July 2003, Debrecen, Hungary
827
In this work, the non- linear function
f
is estimated based on data samples using the
Lavenberg-Marquardt optimisation technique. The Lavenberg-Marquardt is the standard
method for minimization of mean square error criteria, due to its rapid convergence properties
and robustness (Marquardt, 1963). Neural networks have several major drawbacks. They
require large numbers of data samples due to their large number of degrees of freedom.
Problems such as over- fitting and sub-optimal minima may occur more severely than in the
linear case. Also, this method requires a large computation time for training, i.e. for learning
the system behaviour, which restricts its application to real-time implementations.
3. Results and conclusions
The inside air temperature and humidity simulation models were identified using the
described approaches for a greenhouse located in the north of Portugal. The greenhouse has a
floor area of 210m
2
, covered with 200m polyethylene film. Several actuators and sensors are
installed and connected to an acquisition and control system based on a Personal computer
and a data acquisition and control card (PCL-818, from Advantech) using a sampling interval
of 1 minute.
In figure 3, the simulation results achieved with the physical model for the air
temperature and relative humidity are shown. The model parameters were computed off- line
using the data of the month of January 2000, and the simulations were performed for a
validation data set of the first 6 days of March of the same year.
Table 1 shows the performance results of the physical and parametric models for the
pure simulation and the 60 step ahead predictions, which corresponds to predictions in the
future time horizon of 60 minutes.
In this case, the data sets used to compute the models parameters and the simulations
were the same of the previous case and the parametric models were estimated off- line using a
least squares algorithm and a neural network. The criteria performance used is the root mean
squared errors,
2
1
1
n
k
k
RMSE E
N

, with N being the size of the data samples and Ek the
error between simulated and measured values.

EFITA 2003 Conference 5-9. July 2003, Debrecen, Hungary
828


Figure 3. Measured and simulated temperatures (top) and relative humidity (bottom)
Table1 Performance of the air temperature and relative humidity models for the data
set validation
Air temperature Air relative humidity
Model RMSE
(simulation)
RMSE
(60 min
prediction)
RMSE
(simulation)
RMSE
(60 min
prediction)
Physical 1.4063 0.1928 4.9401 0.6914
ARX 2.0022 0.2231 5.2810 0.9933
Neural
Network
1.8309 0.2198 5.1772 0.9614
EFITA 2003 Conference 5-9. July 2003, Debrecen, Hungary
829
This study has drawn several comparisons between physical, linear and non- linear
modelling techniques applied to simulate the inside greenhouse climate. It was observed from
various simulations that non recursive physical models give better results than other off- line
methods, when are used validation data sets periods distant in time from the data sets used to
compute the models. However, computation times for ARX models are much lower, and if
recursive estimation techniques are applied, the results achieved for the short time prediction
horizon, from 1 to 60 minutes, are better with these models. The use of neural networks
models has the major drawback of requiring a large computation time for training, which
restricts their application to real-time implementations.

REFERENCES
Astrm, K. J. And B. Wittenmark (1989). Adaptive control, Addison - Wesley, Massachusetts.
Bakker, J. C., Bot, G.P.A., Challa, H., N.J. van de (1994). Greenhouse climate control, Wageningen Pers
Publishers, Wageningen, 1994, 384pp.
Bot, G.P.A. (1991). Physical modelling of greenhouse climate. Proceedings of the IFAC/ISHS Workshop, pp: 7-
12.
Boulard, T. and A. Baille, (1993). A simple greenhouse climate control model incorporating effects on
ventilation and evaporative cooling. Agricultural and Forest Meteorology, 65, pp:145-157.
Hornik K., Stinchcombe M. and White H. 1989. Multi-Layer Feedforward Networks are Universal
Aproximators, Neural Networks 2, 359-366.
J. Boaventura Cunha., C. Couto and A.E.B. Ruano, (1997). Real-time parameter estimation of dynamic
temperature models for greenhouse environmental control. Control Eng. Practice, Vol. 5, N.10, pp.
1473-1481.
J. Boaventura Cunha, Carlos Couto, A.E.B. Ruano (2000). A greenhouse Climate Multivariable Predictive
Controller, Acta Horticulturae N. 534, ISHS, pp:269-276.
J. P. Coelho, J. Boaventura Cunha, P. B. de Moura Oliveira (2002). Solar radiation Prediction methods applied
to improve greenhouse climate control, World Congress of Computers in Agriculture and Natural
Resources, 13-15 March, 2002, pp:154-161.
Ljung, L. (1987). System identification - theory for the user, PTR Prentice-Hall, New Jersey.
Marquardt, D. 1963. An Algorithm for Least-Squares Estimation of Nonlinear Parameters, SIAM J. Appl. Math.
11, pp. 164-168.
Salgado, Mario E., Graham C. Goodwin and Richard H. Middleton, (1988). Modified least squares algorithm
incorporating exponential resetting and forgetting. Int. J. Control, 21, 477-491